[SHARING] DFL 2.0 ReadyToTrain model

[iperov]

This is heavily pretrained model to train your own Ready-To-Merge model for 224 res using RTM WF Faceset.

How to use this model.

1. place RTM WF Faceset to workspace/data_dst/aligned

2. place your celeb to workspace/data_src/aligned

3. place this model folder to workspace/model

4. do not change settings, train +500.000 iters

5. random_warp OFF, train +500.000, periodically disable masked training for 5.000 iters and enable again

6. GAN 0.1 power, patch size 28, gan_dims:32. Train until the src loss value has not increased in the last 12 hours

7. finalize model by disabling masked training for 100-200 (not thousand) iterations

 

[wangzhang320]

Hi guys, DF architecture can also produce ready-to-merge model，
just turn on true face a little（0.01-0.1） and put FFHQ or celebA dataset in dst.
here's the result

 

[TMBDF]

Hi guys, DF architecture can also produce ready-to-merge model，
just turn on true face a little（0.01-0.1） and put FFHQ or celebA dataset in dst.
here's the result

Using true face will make the lighting look bad, DF architecture already struggles more with lighting, adding TF will only make it worse.

 

[Rfedz007]

-----
Suggest description.
-----


ReadyToMerge model can be used to swap the face without the training.

But you can continue training with the same parameters on the new dst and improve the quality.

----------------
ReadyToMerge LIAE-UD model can be reused to train new fake from scratch.
All you need is delete INTER_B.npy file from the model folder. Also don't forget to set training settings to initial, such as RW(ON), LRD(off), GAN(0.0), etc...

The generalization with poor dst faceset will be better than training with default pretrained model.

Example using Tom Cruise model to train Keanu Reeves fake on poor dst faceset from the game:

----------------

Download: <a href='/literotica/magnet:?xt=urn:btih:dfa9f27340097eb47c0c3cb9e7a8cb6d9876580a&dn=DeepFaceLabReadyToMergeModels&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce'>magnet link</a>

link is broken again. Can you fix it forever?

available celebs:

Tom Cruise (224, liae-ud)
[img=400x276]

I am not able to download ReadyToMerge model. Pls can anyone assist. I am not able to use torrent as it's blocked here

 

[Rfedz007]

-----
Suggest description.
-----


ReadyToMerge model can be used to swap the face without the training.

But you can continue training with the same parameters on the new dst and improve the quality.

----------------
ReadyToMerge LIAE-UD model can be reused to train new fake from scratch.
All you need is delete INTER_B.npy file from the model folder. Also don't forget to set training settings to initial, such as RW(ON), LRD(off), GAN(0.0), etc...

The generalization with poor dst faceset will be better than training with default pretrained model.

Example using Tom Cruise model to train Keanu Reeves fake on poor dst faceset from the game:

----------------

Download: <a href='/literotica/magnet:?xt=urn:btih:dfa9f27340097eb47c0c3cb9e7a8cb6d9876580a&dn=DeepFaceLabReadyToMergeModels&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce'>magnet link</a>

link is broken again. Can you fix it forever?

available celebs:

Tom Cruise (224, liae-ud)
[img=400x276]

I am not able to download ReadyToMerge model. Pls can anyone assist. I am not able to use torrent as it's blocked here

magnet link is not working as it's showing 0% for the past 15 mins when I tried to download.

---

-----
Suggest description.
-----


ReadyToMerge model can be used to swap the face without the training.

But you can continue training with the same parameters on the new dst and improve the quality.

----------------
ReadyToMerge LIAE-UD model can be reused to train new fake from scratch.
All you need is delete INTER_B.npy file from the model folder. Also don't forget to set training settings to initial, such as RW(ON), LRD(off), GAN(0.0), etc...

The generalization with poor dst faceset will be better than training with default pretrained model.

Example using Tom Cruise model to train Keanu Reeves fake on poor dst faceset from the game:

----------------

Download: <a href='/literotica/magnet:?xt=urn:btih:dfa9f27340097eb47c0c3cb9e7a8cb6d9876580a&dn=DeepFaceLabReadyToMergeModels&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce'>magnet link</a>

link is broken again. Can you fix it forever?

available celebs:

Tom Cruise (224, liae-ud)
[img=400x276]

I am not able to download ReadyToMerge model. Pls can anyone assist. I am not able to use torrent as it's blocked here

magnet link is not working as it's showing 0% for the past 15 mins when I tried to download.

Now it's working. Thanks

 

[Rfedz007]

-----
Suggest description.
-----


ReadyToMerge model can be used to swap the face without the training.

But you can continue training with the same parameters on the new dst and improve the quality.

----------------
ReadyToMerge LIAE-UD model can be reused to train new fake from scratch.
All you need is delete INTER_B.npy file from the model folder. Also don't forget to set training settings to initial, such as RW(ON), LRD(off), GAN(0.0), etc...

The generalization with poor dst faceset will be better than training with default pretrained model.

Example using Tom Cruise model to train Keanu Reeves fake on poor dst faceset from the game:

----------------

Download: <a href='/literotica/magnet:?xt=urn:btih:dfa9f27340097eb47c0c3cb9e7a8cb6d9876580a&dn=DeepFaceLabReadyToMergeModels&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce'>magnet link</a>

link is broken again. Can you fix it forever?

available celebs:

Tom Cruise (224, liae-ud)
[img=400x276]

I am not able to download ReadyToMerge model. Pls can anyone assist. I am not able to use torrent as it's blocked here

magnet link is not working as it's showing 0% for the past 15 mins when I tried to download.

---

-----
Suggest description.
-----


ReadyToMerge model can be used to swap the face without the training.

But you can continue training with the same parameters on the new dst and improve the quality.

----------------
ReadyToMerge LIAE-UD model can be reused to train new fake from scratch.
All you need is delete INTER_B.npy file from the model folder. Also don't forget to set training settings to initial, such as RW(ON), LRD(off), GAN(0.0), etc...

The generalization with poor dst faceset will be better than training with default pretrained model.

Example using Tom Cruise model to train Keanu Reeves fake on poor dst faceset from the game:

----------------

Download: <a href='/literotica/magnet:?xt=urn:btih:dfa9f27340097eb47c0c3cb9e7a8cb6d9876580a&dn=DeepFaceLabReadyToMergeModels&tr=udp%3a%2f%2ftracker.opentrackr.org%3a1337%2fannounce'>magnet link</a>

link is broken again. Can you fix it forever?

available celebs:

Tom Cruise (224, liae-ud)
[img=400x276]

I am not able to download ReadyToMerge model. Pls can anyone assist. I am not able to use torrent as it's blocked here

magnet link is not working as it's showing 0% for the past 15 mins when I tried to download.

Now it's working. Thanks

---

My Training model shuts down after every 5 minutes. Below is the mentioned error which I get frequently.

Starting. Press "Enter" to stop training and save model.
Error: [Errno 13] Permission denied: 'D:\\Jan 2021\\NVIDIA Jan 4th\\DeepFaceLab_NVIDIA\\workspace\\model\\new_SAEHD_summary.txt'
Traceback (most recent call last):
  File "D:\Jan 2021\NVIDIA Jan 4th\DeepFaceLab_NVIDIA\_internal\DeepFaceLab\mainscripts\Trainer.py", line 170, in trainerThread
    model_save()
  File "D:\Jan 2021\NVIDIA Jan 4th\DeepFaceLab_NVIDIA\_internal\DeepFaceLab\mainscripts\Trainer.py", line 69, in model_save
    model.save()
  File "D:\Jan 2021\NVIDIA Jan 4th\DeepFaceLab_NVIDIA\_internal\DeepFaceLab\models\ModelBase.py", line 379, in save
    Path( self.get_summary_path() ).write_text( self.get_summary_text() )
  File "pathlib.py", line 1215, in write_text
  File "pathlib.py", line 1183, in open
  File "pathlib.py", line 1037, in _opener
  File "pathlib.py", line 387, in wrapped
PermissionError: [Errno 13] Permission denied: 'D:\\Jan 2021\\NVIDIA Jan 4th\\DeepFaceLab_NVIDIA\\workspace\\model\\new_SAEHD_summary.txt'
Done.
Press any key to continue . . .

Pls anyone Assist me where I am going wrong...
Thanks

 

[testerson34]

I think I sort of did this on accident as I as learning how to do this stuff. The primary reason I learned it was because I wanted to see a specific person in lots of scenes.

I started with all the basic model settings with DF. No pre-training because I didn't know about that. Just started training my SRC against the first DST and let it run for a couple of days.

From there I did that same SRC on about 20 different scenes. I've also done 6 or 7 other SRC's on this same model against a variety of other DST's as well.

From there I realized that if I wanted to do another DST of any of these SRC's, I could turn them around very quickly. My model is now at 5M+ trained iterations on just these few SRC's, and if I want to make a new DST out of any of them it happens very quickly.

Is that essentially the same thing you've done here?

On a similar note, I've been a little confused by a couple of things. I just got my new rig with an RTX 3090, and I wanted to start over the "correct" way. Using the female only faceset available here, with some of my own facesets combined, I've got it pre-training, up to about 400k iterations right now. This one is higher res with DF-UD instead of just DF.

This one isn't working nearly as quickly as my other one. I guess that's simply a matter of only 400k pretrained iterations vs. 5M+ fully trained iterations...?? If I pre-train this thing up to 5M would it perform the same, or still not as well since it wasn't so focused on a single SRC already?

In the guide about pretraining it talks about how this is useful because you can start with the pretrained model, and you "don't have to start from zero every time."

Am I really supposed to be starting over every time I do a new SRC/DST? Again, it seems that by not doing that it actually speeds things up greatly. Is there some disadvantage to using the same model for millions of iterations with a variety of SRC's and DST's?

EDIT: I think this is different than what I explained, but I'm not sure I fully understand what this is. Any clarification would be awesome.

 

[Mr.BaBe]

the files in your model directory are read only, after 15 minutes when it automatically try to save, you get this error. change those files from read only to be able to write, and it will be ok "Rfedz007"

 

[Rfedz007]

the files in your model directory are read only, after 15 minutes when it automatically try to save, you get this error. change those files  from read only to be able to write, and it will be ok "Rfedz007"

Thanks a lot now it's working fine & even auto saves after 15 minutes.

 

[iperov]

-Enabled pretraining, immediately saved and closed the model and disabled pretraining

don't do that. Enable/disable pretraining deletes also inter_AB.

 

[alucardvii]

-Enabled pretraining, immediately saved and closed the model and disabled pretraining

don't do that. Enable/disable pretraining deletes also inter_AB.

Thanks for the reply, Iperov! Good to know. Mad respect for the amazing work you do.

 

[light1up]

Great work! Going to try this out now. Be dope to get this in a higher resolution at some point in the future

 

[alucardvii]

Am I really supposed to be starting over every time I do a new SRC/DST?  Again, it seems that by not doing that it actually speeds things up greatly.  Is there some disadvantage to using the same model for millions of iterations with a variety of SRC's and DST's?

Based on what Tuts said, I'm guessing you really did create something similar to what Iperov posted here. I've noticed the same effect myself. 

And no, you're correct - you don't restart a model or do pretraining once you have an established working model. You re-enable random warp each time you swap DSTs, but that's it. 

After you switched model architectures, yeah, you lost the good training you had established on your previous model. But if you train your new model up to 5M iterations again, you'll probably find the performance becomes comparable to your old model (or even better). 

My gut instincts tell me that fucking around with your SRC is probably bad for your model's learning, though. I'd try to keep SRC as consistent as possible if you're trying to make a model that produces results as fast as it can.

 

[selinadfl]

Oh my god - posting in a potentially legendary thread. Could this be the most revolutionary thing to ever happen to deepfake technology since its inception???

Iperov, this is AMAZING. Thank you so much for sharing your hard work with the world!

What I did:
-Downloaded the model and deleted INTER_B.npy, as Iperov instructed.
-Figured out batch settings that would let the model run on my graphics card (GTX 1060: batch 8, no adabelief).
-Enabled pretraining, immediately saved and closed the model and disabled pretraining (this 'clears out' Tom Cruise and resets the loss graphs. Is there any way to reset the iteration count/graphs of a model without doing this?).
-Began training with my own SRC and DST regularly, as described by Iperov above and the detailed guide written by "Tuts" (RW on, LRD off, GAN 0.0).
-Results so far are FANTASTIC for the amount of training time.

If the model keeps performing the way it's doing now, it will greatly reduce the time necessary to create a deepfake (using the 200k iteration as a guess, maybe by a factor of 0.4).

Hi there where is Tuts detailed guide pls share link...............thx................I want to swap without training

 

[alucardvii]

Hi there where is Tuts detailed guide pls share link...............thx................I want to swap without training

Uh, well, the model Iperov posted in this thread only lets you swap to Tom Cruise without training. 
If you want to swap to someone else, you still need to do at least 200k iterations of training for your first clip. 

//deep.whitecatchel.ru/literotica/forums/thread-guide-deepfacelab-2-0-guide

That's Tuts' guide. If you're new, study it carefully - it has everything you need to know; follow the guide exactly to produce good results.
If you try to take shortcuts or do things different, your result will not be good; you'll have to re-start everything, and you'll wish you had followed the guide precisely in the first place. 
If you ever see someone disagree with the guide, they're probably wrong. 

I personally found that Iperov's model in this thread did train a bit faster than regular pretrained models, though. For a 10min clip, I got okay results with this many iterations in each stage of training:
175k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD (I personally don't like GAN, it tends to give me more problems than benefits). 
I probably could've done more training in each stage if I wanted a perfect fake, though.

 

[selinadfl]

Oh, I see, I am being a member since 2018, but I stopped making deepfake last 10months. I was really excited about direct swap...hehe.

 

[iperov]

updated How to create RTM model from scratch.

 

[light1up]

Hi there where is Tuts detailed guide pls share link...............thx................I want to swap without training

Uh, well, the model Iperov posted in this thread only lets you swap to Tom Cruise without training. 
If you want to swap to someone else, you still need to do at least 200k iterations of training for your first clip. 

//deep.whitecatchel.ru/literotica/forums/thread-guide-deepfacelab-2-0-guide

That's Tuts' guide. If you're new, study it carefully - it has everything you need to know; follow the guide exactly to produce good results.
If you try to take shortcuts or do things different, your result will not be good; you'll have to re-start everything, and you'll wish you had followed the guide precisely in the first place. 
If you ever see someone disagree with the guide, they're probably wrong. 

I personally found that Iperov's model in this thread did train a bit faster than regular pretrained models, though. For a 10min clip, I got okay results with this many iterations in each stage of training:
1750k RW -> 60k RW LRD -> 35k no-RW -> 20k no-RW LRD (I personally don't like GAN, it tends to give me more problems than benefits). 
I probably could've done more training in each stage if I wanted a perfect fake, though.

Is that a typo or did you mean 1 million and 750 thousand iterations? Is that ontop of iperov's 1.3 million or in total including that? I usually do few hundred thousand and get good results but maybe I should start leaving it for longer at the RW stage.

 

[alucardvii]

Is that a typo or did you mean 1 million and 750 thousand iterations? Is that ontop of iperov's 1.3 million or in total including that? I usually do few hundred thousand and get good results but maybe I should start leaving it for longer at the RW stage.

Oops, yes, sorry, 175k. 
And yeah, that's all on top of Iperov's training. 

As I understand it, when your loss values remain stable across a few hours, there's little benefit to doing any more training during a given stage, and you can proceed to the next stage.

 

[wangzhang320]

Hi guys, DF architecture can also produce ready-to-merge model，
just turn on true face a little（0.01-0.1） and put FFHQ or celebA dataset in dst.
here's the result

password: dfldata.xyz